Graph Neural Network-Based Simulation of Ocean-Atmosphere Coupling Processes and Their Impact on Tropical Cyclone Intensity Prediction

Abstract:

Ocean-atmosphere coupling processes are core components of the Earth system, regulating global energy redistribution and regional extreme weather events, especially tropical cyclones (TCs). Traditional coupled ocean-atmosphere models rely on grid-based parameterization schemes to describe nonlinear coupling processes, leading to significant uncertainties in simulating key variables such as sea surface temperature (SST) anomalies, latent heat flux (LHF) exchange, and thus TC intensity prediction. This study proposes a Graph Neural Network-based ocean-atmosphere coupling simulation framework (GNN-OAC) that constructs a spatial-temporal graph structure based on the physical connections between ocean and atmosphere grid cells to capture the non-local and nonlinear coupling relationships. The framework assimilates multi-source observation data, including satellite-derived SST, LHF, and in-situ buoy data, to optimize the representation of key coupling processes such as ocean mixed layer heat exchange and atmospheric boundary layer instability. Validation results based on 30 years of global TC observation data show that the GNN-OAC framework improves the average accuracy of TC intensity prediction by 23% compared to the traditional Coupled Ocean-Atmosphere Response Experiment (COARE) model. Simulation under CMIP6 SSP2-4.5 and SSP5-8.5 scenarios indicates that the GNN-OAC framework reduces the uncertainty of TC intensity projection by 18-22% by the end of the 21st century. Specifically, in the western North Pacific and North Atlantic TC basins, the framework accurately captures the negative feedback effect of SST cooling on TC intensity enhancement and the positive feedback effect of atmospheric convection on ocean upwelling. This study provides a new paradigm for improving the simulation accuracy of ocean-atmosphere coupling processes in Earth system models and enhances the reliability of extreme weather event prediction, offering important scientific support for formulating marine disaster prevention and mitigation strategies.